From f1625f274acf5dcd5601f6cb5e29e233b2a441a3 Mon Sep 17 00:00:00 2001
From: djm <>
Date: Sat, 6 Sep 2008 12:15:54 +0000
Subject: import of OpenSSL 0.9.8h

---
 src/lib/libcrypto/perlasm/x86ms.pl   | 122 ++++++++++++++++---
 src/lib/libcrypto/perlasm/x86nasm.pl | 153 ++++++++++++++++++------
 src/lib/libcrypto/perlasm/x86unix.pl | 220 ++++++++++++++++++++++++++++-------
 3 files changed, 404 insertions(+), 91 deletions(-)

(limited to 'src/lib/libcrypto/perlasm')

diff --git a/src/lib/libcrypto/perlasm/x86ms.pl b/src/lib/libcrypto/perlasm/x86ms.pl
index b6bd744057..a0be2934c2 100644
--- a/src/lib/libcrypto/perlasm/x86ms.pl
+++ b/src/lib/libcrypto/perlasm/x86ms.pl
@@ -27,7 +27,13 @@ $label="L000";
 sub main'asm_init_output { @out=(); }
 sub main'asm_get_output { return(@out); }
 sub main'get_labels { return(@labels); }
-sub main'external_label { push(@labels,@_); }
+sub main'external_label
+{
+	push(@labels,@_);
+	foreach (@_) {
+		push(@out, "EXTRN\t_$_:DWORD\n");
+	}
+}
 
 sub main'LB
 	{
@@ -51,6 +57,11 @@ sub main'DWP
 	&get_mem("DWORD",@_);
 	}
 
+sub main'QWP
+	{
+	&get_mem("QWORD",@_);
+	}
+
 sub main'BC
 	{
 	return @_;
@@ -87,7 +98,7 @@ sub get_mem
 		$reg2=&conv($1);
 		$addr="_$2";
 		}
-	elsif ($addr =~ /^[_a-zA-Z]/)
+	elsif ($addr =~ /^[_a-z][_a-z0-9]*$/i)
 		{
 		$addr="_$addr";
 		}
@@ -128,12 +139,14 @@ sub main'xorb	{ &out2("xor",@_); }
 sub main'add	{ &out2("add",@_); }
 sub main'adc	{ &out2("adc",@_); }
 sub main'sub	{ &out2("sub",@_); }
+sub main'sbb	{ &out2("sbb",@_); }
 sub main'rotl	{ &out2("rol",@_); }
 sub main'rotr	{ &out2("ror",@_); }
 sub main'exch	{ &out2("xchg",@_); }
 sub main'cmp	{ &out2("cmp",@_); }
 sub main'lea	{ &out2("lea",@_); }
 sub main'mul	{ &out1("mul",@_); }
+sub main'imul	{ &out2("imul",@_); }
 sub main'div	{ &out1("div",@_); }
 sub main'dec	{ &out1("dec",@_); }
 sub main'inc	{ &out1("inc",@_); }
@@ -155,26 +168,54 @@ sub main'jne	{ &out1("jne",@_); }
 sub main'jno	{ &out1("jno",@_); }
 sub main'push	{ &out1("push",@_); $stack+=4; }
 sub main'pop	{ &out1("pop",@_); $stack-=4; }
+sub main'pushf	{ &out0("pushfd"); $stack+=4; }
+sub main'popf	{ &out0("popfd"); $stack-=4; }
 sub main'bswap	{ &out1("bswap",@_); &using486(); }
 sub main'not	{ &out1("not",@_); }
 sub main'call	{ &out1("call",($_[0]=~/^\$L/?'':'_').$_[0]); }
+sub main'call_ptr { &out1p("call",@_); }
 sub main'ret	{ &out0("ret"); }
 sub main'nop	{ &out0("nop"); }
+sub main'test	{ &out2("test",@_); }
+sub main'bt	{ &out2("bt",@_); }
+sub main'leave	{ &out0("leave"); }
+sub main'cpuid  { &out0("DW\t0A20Fh"); }
+sub main'rdtsc  { &out0("DW\t0310Fh"); }
+sub main'halt	{ &out0("hlt"); }
 sub main'movz	{ &out2("movzx",@_); }
+sub main'neg	{ &out1("neg",@_); }
+sub main'cld	{ &out0("cld"); }
+
+# SSE2
+sub main'emms	{ &out0("emms"); }
+sub main'movd	{ &out2("movd",@_); }
+sub main'movq	{ &out2("movq",@_); }
+sub main'movdqu	{ &out2("movdqu",@_); }
+sub main'movdqa	{ &out2("movdqa",@_); }
+sub main'movdq2q{ &out2("movdq2q",@_); }
+sub main'movq2dq{ &out2("movq2dq",@_); }
+sub main'paddq	{ &out2("paddq",@_); }
+sub main'pmuludq{ &out2("pmuludq",@_); }
+sub main'psrlq	{ &out2("psrlq",@_); }
+sub main'psllq	{ &out2("psllq",@_); }
+sub main'pxor	{ &out2("pxor",@_); }
+sub main'por	{ &out2("por",@_); }
+sub main'pand	{ &out2("pand",@_); }
 
 sub out2
 	{
 	local($name,$p1,$p2)=@_;
-	local($l,$t);
+	local($l,$t,$line);
 
-	push(@out,"\t$name\t");
+	$line="\t$name\t";
 	$t=&conv($p1).",";
 	$l=length($t);
-	push(@out,$t);
+	$line.="$t";
 	$l=4-($l+9)/8;
-	push(@out,"\t" x $l);
-	push(@out,&conv($p2));
-	push(@out,"\n");
+	$line.="\t" x $l;
+	$line.=&conv($p2);
+	if ($line=~/\bxmm[0-7]\b/i) { $line=~s/\b[A-Z]+WORD\s+PTR/XMMWORD PTR/i; }
+	push(@out,$line."\n");
 	}
 
 sub out0
@@ -214,7 +255,9 @@ sub main'file
 	local($tmp)=<<"EOF";
 	TITLE	$file.asm
         .386
-.model FLAT
+.model	FLAT
+_TEXT\$	SEGMENT PAGE 'CODE'
+
 EOF
 	push(@out,$tmp);
 	}
@@ -226,7 +269,6 @@ sub main'function_begin
 	push(@labels,$func);
 
 	local($tmp)=<<"EOF";
-_TEXT	SEGMENT
 PUBLIC	_$func
 $extra
 _$func PROC NEAR
@@ -244,7 +286,6 @@ sub main'function_begin_B
 	local($func,$extra)=@_;
 
 	local($tmp)=<<"EOF";
-_TEXT	SEGMENT
 PUBLIC	_$func
 $extra
 _$func PROC NEAR
@@ -264,7 +305,6 @@ sub main'function_end
 	pop	ebp
 	ret
 _$func ENDP
-_TEXT	ENDS
 EOF
 	push(@out,$tmp);
 	$stack=0;
@@ -277,7 +317,6 @@ sub main'function_end_B
 
 	local($tmp)=<<"EOF";
 _$func ENDP
-_TEXT	ENDS
 EOF
 	push(@out,$tmp);
 	$stack=0;
@@ -300,6 +339,20 @@ EOF
 
 sub main'file_end
 	{
+	# try to detect if SSE2 or MMX extensions were used...
+	my $xmmheader=<<___;
+.686
+.XMM
+IF \@Version LT 800
+XMMWORD STRUCT 16
+	DQ  2 dup (?)
+XMMWORD ENDS
+ENDIF
+___
+	if (grep {/\b[x]?mm[0-7]\b/i} @out) {
+		grep {s/\.[3-7]86/$xmmheader/} @out;
+		}
+	push(@out,"_TEXT\$	ENDS\n");
 	push(@out,"END\n");
 	}
 
@@ -331,6 +384,12 @@ sub main'comment
 		}
 	}
 
+sub main'public_label
+	{
+	$label{$_[0]}="_$_[0]"	if (!defined($label{$_[0]}));
+	push(@out,"PUBLIC\t$label{$_[0]}\n");
+	}
+
 sub main'label
 	{
 	if (!defined($label{$_[0]}))
@@ -348,19 +407,37 @@ sub main'set_label
 		$label{$_[0]}="\$${label}${_[0]}";
 		$label++;
 		}
+	if ($_[1]!=0 && $_[1]>1)
+		{
+		main'align($_[1]);
+		}
 	if((defined $_[2]) && ($_[2] == 1))
 		{
 		push(@out,"$label{$_[0]}::\n");
 		}
+	elsif ($label{$_[0]} !~ /^\$/)
+		{
+		push(@out,"$label{$_[0]}\tLABEL PTR\n");
+		}
 	else
 		{
 		push(@out,"$label{$_[0]}:\n");
 		}
 	}
 
+sub main'data_byte
+	{
+	push(@out,"\tDB\t".join(',',@_)."\n");
+	}
+
 sub main'data_word
 	{
-	push(@out,"\tDD\t$_[0]\n");
+	push(@out,"\tDD\t".join(',',@_)."\n");
+	}
+
+sub main'align
+	{
+	push(@out,"\tALIGN\t$_[0]\n");
 	}
 
 sub out1p
@@ -368,7 +445,7 @@ sub out1p
 	local($name,$p1)=@_;
 	local($l,$t);
 
-	push(@out,"\t$name\t ".&conv($p1)."\n");
+	push(@out,"\t$name\t".&conv($p1)."\n");
 	}
 
 sub main'picmeup
@@ -378,3 +455,18 @@ sub main'picmeup
 	}
 
 sub main'blindpop { &out1("pop",@_); }
+
+sub main'initseg 
+	{
+	local($f)=@_;
+	local($tmp)=<<___;
+OPTION	DOTNAME
+.CRT\$XCU	SEGMENT DWORD PUBLIC 'DATA'
+EXTRN	_$f:NEAR
+DD	_$f
+.CRT\$XCU	ENDS
+___
+	push(@out,$tmp);
+	}
+
+1;
diff --git a/src/lib/libcrypto/perlasm/x86nasm.pl b/src/lib/libcrypto/perlasm/x86nasm.pl
index 4bdb3fe180..fa38f89c09 100644
--- a/src/lib/libcrypto/perlasm/x86nasm.pl
+++ b/src/lib/libcrypto/perlasm/x86nasm.pl
@@ -3,6 +3,7 @@
 package x86nasm;
 
 $label="L000";
+$under=($main'netware)?'':'_';
 
 %lb=(	'eax',	'al',
 	'ebx',	'bl',
@@ -32,7 +33,8 @@ sub main'external_label
 {
 	push(@labels,@_);
 	foreach (@_) {
-		push(@out, "extern\t_$_\n");
+		push(@out,".") if ($main'mwerks);
+		push(@out, "extern\t${under}$_\n");
 	}
 }
 
@@ -58,14 +60,19 @@ sub main'DWP
 	&get_mem("DWORD",@_);
 	}
 
+sub main'QWP
+	{
+	&get_mem("",@_);
+	}
+
 sub main'BC
 	{
-	return "BYTE @_";
+	return (($main'mwerks)?"":"BYTE ")."@_";
 	}
 
 sub main'DWC
 	{
-	return "DWORD @_";
+	return (($main'mwerks)?"":"DWORD ")."@_";
 	}
 
 sub main'stack_push
@@ -86,16 +93,22 @@ sub get_mem
 	{
 	my($size,$addr,$reg1,$reg2,$idx)=@_;
 	my($t,$post);
-	my($ret)="$size [";
+	my($ret)=$size;
+	if ($ret ne "")
+		{
+		$ret .= " PTR" if ($main'mwerks);
+		$ret .= " ";
+		}
+	$ret .= "[";
 	$addr =~ s/^\s+//;
 	if ($addr =~ /^(.+)\+(.+)$/)
 		{
 		$reg2=&conv($1);
-		$addr="_$2";
+		$addr="$under$2";
 		}
-	elsif ($addr =~ /^[_a-zA-Z]/)
+	elsif ($addr =~ /^[_a-z][_a-z0-9]*$/i)
 		{
-		$addr="_$addr";
+		$addr="$under$addr";
 		}
 
 	if ($addr =~ /^.+\-.+$/) { $addr="($addr)"; }
@@ -134,12 +147,14 @@ sub main'xorb	{ &out2("xor",@_); }
 sub main'add	{ &out2("add",@_); }
 sub main'adc	{ &out2("adc",@_); }
 sub main'sub	{ &out2("sub",@_); }
+sub main'sbb	{ &out2("sbb",@_); }
 sub main'rotl	{ &out2("rol",@_); }
 sub main'rotr	{ &out2("ror",@_); }
 sub main'exch	{ &out2("xchg",@_); }
 sub main'cmp	{ &out2("cmp",@_); }
 sub main'lea	{ &out2("lea",@_); }
 sub main'mul	{ &out1("mul",@_); }
+sub main'imul	{ &out2("imul",@_); }
 sub main'div	{ &out1("div",@_); }
 sub main'dec	{ &out1("dec",@_); }
 sub main'inc	{ &out1("inc",@_); }
@@ -147,29 +162,57 @@ sub main'jmp	{ &out1("jmp",@_); }
 sub main'jmp_ptr { &out1p("jmp",@_); }
 
 # This is a bit of a kludge: declare all branches as NEAR.
-sub main'je	{ &out1("je NEAR",@_); }
-sub main'jle	{ &out1("jle NEAR",@_); }
-sub main'jz	{ &out1("jz NEAR",@_); }
-sub main'jge	{ &out1("jge NEAR",@_); }
-sub main'jl	{ &out1("jl NEAR",@_); }
-sub main'ja	{ &out1("ja NEAR",@_); }
-sub main'jae	{ &out1("jae NEAR",@_); }
-sub main'jb	{ &out1("jb NEAR",@_); }
-sub main'jbe	{ &out1("jbe NEAR",@_); }
-sub main'jc	{ &out1("jc NEAR",@_); }
-sub main'jnc	{ &out1("jnc NEAR",@_); }
-sub main'jnz	{ &out1("jnz NEAR",@_); }
-sub main'jne	{ &out1("jne NEAR",@_); }
-sub main'jno	{ &out1("jno NEAR",@_); }
+$near=($main'mwerks)?'':'NEAR';
+sub main'je	{ &out1("je $near",@_); }
+sub main'jle	{ &out1("jle $near",@_); }
+sub main'jz	{ &out1("jz $near",@_); }
+sub main'jge	{ &out1("jge $near",@_); }
+sub main'jl	{ &out1("jl $near",@_); }
+sub main'ja	{ &out1("ja $near",@_); }
+sub main'jae	{ &out1("jae $near",@_); }
+sub main'jb	{ &out1("jb $near",@_); }
+sub main'jbe	{ &out1("jbe $near",@_); }
+sub main'jc	{ &out1("jc $near",@_); }
+sub main'jnc	{ &out1("jnc $near",@_); }
+sub main'jnz	{ &out1("jnz $near",@_); }
+sub main'jne	{ &out1("jne $near",@_); }
+sub main'jno	{ &out1("jno $near",@_); }
 
 sub main'push	{ &out1("push",@_); $stack+=4; }
 sub main'pop	{ &out1("pop",@_); $stack-=4; }
+sub main'pushf	{ &out0("pushfd"); $stack+=4; }
+sub main'popf	{ &out0("popfd"); $stack-=4; }
 sub main'bswap	{ &out1("bswap",@_); &using486(); }
 sub main'not	{ &out1("not",@_); }
-sub main'call	{ &out1("call",($_[0]=~/^\$L/?'':'_').$_[0]); }
+sub main'call	{ &out1("call",($_[0]=~/^\@L/?'':$under).$_[0]); }
+sub main'call_ptr { &out1p("call",@_); }
 sub main'ret	{ &out0("ret"); }
 sub main'nop	{ &out0("nop"); }
+sub main'test	{ &out2("test",@_); }
+sub main'bt	{ &out2("bt",@_); }
+sub main'leave	{ &out0("leave"); }
+sub main'cpuid	{ &out0("cpuid"); }
+sub main'rdtsc	{ &out0("rdtsc"); }
+sub main'halt	{ &out0("hlt"); }
 sub main'movz	{ &out2("movzx",@_); }
+sub main'neg	{ &out1("neg",@_); }
+sub main'cld	{ &out0("cld"); }
+
+# SSE2
+sub main'emms	{ &out0("emms"); }
+sub main'movd	{ &out2("movd",@_); }
+sub main'movq	{ &out2("movq",@_); }
+sub main'movdqu	{ &out2("movdqu",@_); }
+sub main'movdqa	{ &out2("movdqa",@_); }
+sub main'movdq2q{ &out2("movdq2q",@_); }
+sub main'movq2dq{ &out2("movq2dq",@_); }
+sub main'paddq	{ &out2("paddq",@_); }
+sub main'pmuludq{ &out2("pmuludq",@_); }
+sub main'psrlq	{ &out2("psrlq",@_); }
+sub main'psllq	{ &out2("psllq",@_); }
+sub main'pxor	{ &out2("pxor",@_); }
+sub main'por	{ &out2("por",@_); }
+sub main'pand	{ &out2("pand",@_); }
 
 sub out2
 	{
@@ -177,7 +220,7 @@ sub out2
 	my($l,$t);
 
 	push(@out,"\t$name\t");
-	if ($name eq "lea")
+	if (!$main'mwerks and $name eq "lea")
 		{
 		$p1 =~ s/^[^\[]*\[/\[/;
 		$p2 =~ s/^[^\[]*\[/\[/;
@@ -221,15 +264,17 @@ sub using486
 
 sub main'file
 	{
-	local $tmp;
-	$tmp=<<___;
+	if ($main'mwerks)	{ push(@out,".section\t.text\n"); }
+	else	{
+		local $tmp=<<___;
 %ifdef __omf__
 section	code	use32 class=code
 %else
 section	.text
 %endif
 ___
-	push(@out,$tmp);
+		push(@out,$tmp);
+		}
 	}
 
 sub main'function_begin
@@ -237,9 +282,10 @@ sub main'function_begin
 	my($func,$extra)=@_;
 
 	push(@labels,$func);
+	push(@out,".") if ($main'mwerks);
 	my($tmp)=<<"EOF";
-global	_$func
-_$func:
+global	$under$func
+$under$func:
 	push	ebp
 	push	ebx
 	push	esi
@@ -252,9 +298,10 @@ EOF
 sub main'function_begin_B
 	{
 	my($func,$extra)=@_;
+	push(@out,".") if ($main'mwerks);
 	my($tmp)=<<"EOF";
-global	_$func
-_$func:
+global	$under$func
+$under$func:
 EOF
 	push(@out,$tmp);
 	$stack=4;
@@ -328,11 +375,18 @@ sub main'comment
 		}
 	}
 
+sub main'public_label
+	{
+	$label{$_[0]}="${under}${_[0]}"	if (!defined($label{$_[0]}));
+	push(@out,".") if ($main'mwerks);
+	push(@out,"global\t$label{$_[0]}\n");
+	}
+
 sub main'label
 	{
 	if (!defined($label{$_[0]}))
 		{
-		$label{$_[0]}="\$${label}${_[0]}";
+		$label{$_[0]}="\@${label}${_[0]}";
 		$label++;
 		}
 	return($label{$_[0]});
@@ -342,15 +396,30 @@ sub main'set_label
 	{
 	if (!defined($label{$_[0]}))
 		{
-		$label{$_[0]}="\$${label}${_[0]}";
+		$label{$_[0]}="\@${label}${_[0]}";
 		$label++;
 		}
+	if ($_[1]!=0 && $_[1]>1)
+		{
+		main'align($_[1]);
+		}
 	push(@out,"$label{$_[0]}:\n");
 	}
 
+sub main'data_byte
+	{
+	push(@out,(($main'mwerks)?".byte\t":"DB\t").join(',',@_)."\n");
+	}
+
 sub main'data_word
 	{
-	push(@out,"\tDD\t$_[0]\n");
+	push(@out,(($main'mwerks)?".long\t":"DD\t").join(',',@_)."\n");
+	}
+
+sub main'align
+	{
+	push(@out,".") if ($main'mwerks);
+	push(@out,"align\t$_[0]\n");
 	}
 
 sub out1p
@@ -358,7 +427,7 @@ sub out1p
 	my($name,$p1)=@_;
 	my($l,$t);
 
-	push(@out,"\t$name\t ".&conv($p1)."\n");
+	push(@out,"\t$name\t".&conv($p1)."\n");
 	}
 
 sub main'picmeup
@@ -368,3 +437,19 @@ sub main'picmeup
 	}
 
 sub main'blindpop { &out1("pop",@_); }
+
+sub main'initseg
+	{
+	local($f)=@_;
+	if ($main'win32)
+		{
+		local($tmp)=<<___;
+segment	.CRT\$XCU data
+extern	$under$f
+DD	$under$f
+___
+		push(@out,$tmp);
+		}
+	}
+
+1;
diff --git a/src/lib/libcrypto/perlasm/x86unix.pl b/src/lib/libcrypto/perlasm/x86unix.pl
index a31a25c12b..a4c947165e 100644
--- a/src/lib/libcrypto/perlasm/x86unix.pl
+++ b/src/lib/libcrypto/perlasm/x86unix.pl
@@ -1,14 +1,15 @@
 #!/usr/local/bin/perl
 
-package x86unix;
+package x86unix;	# GAS actually...
 
 $label="L000";
 $const="";
 $constl=0;
 
 $align=($main'aout)?"4":"16";
-$under=($main'aout)?"_":"";
-$com_start=($main'sol)?"/":"#";
+$under=($main'aout or $main'coff)?"_":"";
+$dot=($main'aout)?"":".";
+$com_start="#" if ($main'aout or $main'coff);
 
 sub main'asm_init_output { @out=(); }
 sub main'asm_get_output { return(@out); }
@@ -51,6 +52,24 @@ if ($main'cpp)
 	'edi',	'%edi',
 	'ebp',	'%ebp',
 	'esp',	'%esp',
+
+	'mm0',	'%mm0',
+	'mm1',	'%mm1',
+	'mm2',	'%mm2',
+	'mm3',	'%mm3',
+	'mm4',	'%mm4',
+	'mm5',	'%mm5',
+	'mm6',	'%mm6',
+	'mm7',	'%mm7',
+
+	'xmm0',	'%xmm0',
+	'xmm1',	'%xmm1',
+	'xmm2',	'%xmm2',
+	'xmm3',	'%xmm3',
+	'xmm4',	'%xmm4',
+	'xmm5',	'%xmm5',
+	'xmm6',	'%xmm6',
+	'xmm7',	'%xmm7',
 	);
 
 %reg_val=(
@@ -97,6 +116,11 @@ sub main'DWP
 	return($ret);
 	}
 
+sub main'QWP
+	{
+	return(&main'DWP(@_));
+	}
+
 sub main'BP
 	{
 	return(&main'DWP(@_));
@@ -140,12 +164,14 @@ sub main'xorb	{ &out2("xorb",@_); }
 sub main'add	{ &out2($_[0]=~/%[a-d][lh]/?"addb":"addl",@_); }
 sub main'adc	{ &out2("adcl",@_); }
 sub main'sub	{ &out2("subl",@_); }
+sub main'sbb	{ &out2("sbbl",@_); }
 sub main'rotl	{ &out2("roll",@_); }
 sub main'rotr	{ &out2("rorl",@_); }
 sub main'exch	{ &out2($_[0]=~/%[a-d][lh]/?"xchgb":"xchgl",@_); }
 sub main'cmp	{ &out2("cmpl",@_); }
 sub main'lea	{ &out2("leal",@_); }
 sub main'mul	{ &out1("mull",@_); }
+sub main'imul	{ &out2("imull",@_); }
 sub main'div	{ &out1("divl",@_); }
 sub main'jmp	{ &out1("jmp",@_); }
 sub main'jmp_ptr { &out1p("jmp",@_); }
@@ -167,13 +193,48 @@ sub main'dec	{ &out1("decl",@_); }
 sub main'inc	{ &out1($_[0]=~/%[a-d][hl]/?"incb":"incl",@_); }
 sub main'push	{ &out1("pushl",@_); $stack+=4; }
 sub main'pop	{ &out1("popl",@_); $stack-=4; }
-sub main'pushf	{ &out0("pushf"); $stack+=4; }
-sub main'popf	{ &out0("popf"); $stack-=4; }
+sub main'pushf	{ &out0("pushfl"); $stack+=4; }
+sub main'popf	{ &out0("popfl"); $stack-=4; }
 sub main'not	{ &out1("notl",@_); }
-sub main'call	{ &out1("call",($_[0]=~/^\.L/?'':$under).$_[0]); }
+sub main'call	{	my $pre=$under;
+			foreach $i (%label)
+			{ if ($label{$i} eq $_[0]) { $pre=''; last; } }
+			&out1("call",$pre.$_[0]);
+		}
+sub main'call_ptr { &out1p("call",@_); }
 sub main'ret	{ &out0("ret"); }
 sub main'nop	{ &out0("nop"); }
+sub main'test	{ &out2("testl",@_); }
+sub main'bt	{ &out2("btl",@_); }
+sub main'leave	{ &out0("leave"); }
+sub main'cpuid	{ &out0(".byte\t0x0f,0xa2"); }
+sub main'rdtsc	{ &out0(".byte\t0x0f,0x31"); }
+sub main'halt	{ &out0("hlt"); }
 sub main'movz	{ &out2("movzbl",@_); }
+sub main'neg	{ &out1("negl",@_); }
+sub main'cld	{ &out0("cld"); }
+
+# SSE2
+sub main'emms	{ &out0("emms"); }
+sub main'movd	{ &out2("movd",@_); }
+sub main'movdqu	{ &out2("movdqu",@_); }
+sub main'movdqa	{ &out2("movdqa",@_); }
+sub main'movdq2q{ &out2("movdq2q",@_); }
+sub main'movq2dq{ &out2("movq2dq",@_); }
+sub main'paddq	{ &out2("paddq",@_); }
+sub main'pmuludq{ &out2("pmuludq",@_); }
+sub main'psrlq	{ &out2("psrlq",@_); }
+sub main'psllq	{ &out2("psllq",@_); }
+sub main'pxor	{ &out2("pxor",@_); }
+sub main'por	{ &out2("por",@_); }
+sub main'pand	{ &out2("pand",@_); }
+sub main'movq	{
+	local($p1,$p2,$optimize)=@_;
+	if ($optimize && $p1=~/^mm[0-7]$/ && $p2=~/^mm[0-7]$/)
+		# movq between mmx registers can sink Intel CPUs
+		{	push(@out,"\tpshufw\t\$0xe4,%$p2,%$p1\n");	}
+	else	{	&out2("movq",@_);				}
+	}
 
 # The bswapl instruction is new for the 486. Emulate if i386.
 sub main'bswap
@@ -279,8 +340,6 @@ sub main'file
 
 	local($tmp)=<<"EOF";
 	.file	"$file.s"
-	.version	"01.01"
-gcc2_compiled.:
 EOF
 	push(@out,$tmp);
 	}
@@ -294,15 +353,17 @@ sub main'function_begin
 
 	local($tmp)=<<"EOF";
 .text
-	.align $align
-.globl $func
+.globl	$func
 EOF
 	push(@out,$tmp);
 	if ($main'cpp)
-		{ $tmp=push(@out,"\tTYPE($func,\@function)\n"); }
-	elsif ($main'gaswin)
-		{ $tmp=push(@out,"\t.def\t$func;\t.scl\t2;\t.type\t32;\t.endef\n"); }
-	else	{ $tmp=push(@out,"\t.type\t$func,\@function\n"); }
+		{ $tmp=push(@out,"TYPE($func,\@function)\n"); }
+	elsif ($main'coff)
+		{ $tmp=push(@out,".def\t$func;\t.scl\t2;\t.type\t32;\t.endef\n"); }
+	elsif ($main'aout and !$main'pic)
+		{ }
+	else	{ $tmp=push(@out,".type\t$func,\@function\n"); }
+	push(@out,".align\t$align\n");
 	push(@out,"$func:\n");
 	$tmp=<<"EOF";
 	pushl	%ebp
@@ -324,15 +385,17 @@ sub main'function_begin_B
 
 	local($tmp)=<<"EOF";
 .text
-	.align $align
-.globl $func
+.globl	$func
 EOF
 	push(@out,$tmp);
 	if ($main'cpp)
-		{ push(@out,"\tTYPE($func,\@function)\n"); }
-	elsif ($main'gaswin)
-		{ $tmp=push(@out,"\t.def\t$func;\t.scl\t2;\t.type\t32;\t.endef\n"); }
-	else	{ push(@out,"\t.type	$func,\@function\n"); }
+		{ push(@out,"TYPE($func,\@function)\n"); }
+	elsif ($main'coff)
+		{ $tmp=push(@out,".def\t$func;\t.scl\t2;\t.type\t32;\t.endef\n"); }
+	elsif ($main'aout and !$main'pic)
+		{ }
+	else	{ push(@out,".type	$func,\@function\n"); }
+	push(@out,".align\t$align\n");
 	push(@out,"$func:\n");
 	$stack=4;
 	}
@@ -349,15 +412,15 @@ sub main'function_end
 	popl	%ebx
 	popl	%ebp
 	ret
-.L_${func}_end:
+${dot}L_${func}_end:
 EOF
 	push(@out,$tmp);
 
 	if ($main'cpp)
-		{ push(@out,"\tSIZE($func,.L_${func}_end-$func)\n"); }
-	elsif ($main'gaswin)
-                { $tmp=push(@out,"\t.align 4\n"); }
-	else	{ push(@out,"\t.size\t$func,.L_${func}_end-$func\n"); }
+		{ push(@out,"SIZE($func,${dot}L_${func}_end-$func)\n"); }
+	elsif ($main'coff or $main'aout)
+                { }
+	else	{ push(@out,".size\t$func,${dot}L_${func}_end-$func\n"); }
 	push(@out,".ident	\"$func\"\n");
 	$stack=0;
 	%label=();
@@ -383,13 +446,13 @@ sub main'function_end_B
 
 	$func=$under.$func;
 
-	push(@out,".L_${func}_end:\n");
+	push(@out,"${dot}L_${func}_end:\n");
 	if ($main'cpp)
-		{ push(@out,"\tSIZE($func,.L_${func}_end-$func)\n"); }
-        elsif ($main'gaswin)
-                { push(@out,"\t.align 4\n"); }
-	else	{ push(@out,"\t.size\t$func,.L_${func}_end-$func\n"); }
-	push(@out,".ident	\"desasm.pl\"\n");
+		{ push(@out,"SIZE($func,${dot}L_${func}_end-$func)\n"); }
+        elsif ($main'coff or $main'aout)
+                { }
+	else	{ push(@out,".size\t$func,${dot}L_${func}_end-$func\n"); }
+	push(@out,".ident	\"$func\"\n");
 	$stack=0;
 	%label=();
 	}
@@ -430,9 +493,10 @@ sub main'swtmp
 
 sub main'comment
 	{
-	if ($main'elf)	# GNU and SVR4 as'es use different comment delimiters,
-		{	# so we just skip comments...
-		push(@out,"\n");
+	if (!defined($com_start) or $main'elf)
+		{	# Regarding $main'elf above...
+			# GNU and SVR4 as'es use different comment delimiters,
+		push(@out,"\n");	# so we just skip ELF comments...
 		return;
 		}
 	foreach (@_)
@@ -444,11 +508,17 @@ sub main'comment
 		}
 	}
 
+sub main'public_label
+	{
+	$label{$_[0]}="${under}${_[0]}"	if (!defined($label{$_[0]}));
+	push(@out,".globl\t$label{$_[0]}\n");
+	}
+
 sub main'label
 	{
 	if (!defined($label{$_[0]}))
 		{
-		$label{$_[0]}=".${label}${_[0]}";
+		$label{$_[0]}="${dot}${label}${_[0]}";
 		$label++;
 		}
 	return($label{$_[0]});
@@ -458,15 +528,29 @@ sub main'set_label
 	{
 	if (!defined($label{$_[0]}))
 		{
-		$label{$_[0]}=".${label}${_[0]}";
+		$label{$_[0]}="${dot}${label}${_[0]}";
 		$label++;
 		}
-	push(@out,".align $align\n") if ($_[1] != 0);
+	if ($_[1]!=0)
+		{
+		if ($_[1]>1)	{ main'align($_[1]);		}
+		else		{ push(@out,".align $align\n");	}
+		}
 	push(@out,"$label{$_[0]}:\n");
 	}
 
 sub main'file_end
 	{
+	# try to detect if SSE2 or MMX extensions were used on ELF platform...
+	if ($main'elf && grep {/\b%[x]*mm[0-7]\b|OPENSSL_ia32cap_P\b/i} @out) {
+		local($tmp);
+
+		push (@out,"\n.section\t.bss\n");
+		push (@out,".comm\t${under}OPENSSL_ia32cap_P,4,4\n");
+
+		return;
+	}
+
 	if ($const ne "")
 		{
 		push(@out,".section .rodata\n");
@@ -475,9 +559,25 @@ sub main'file_end
 		}
 	}
 
+sub main'data_byte
+	{
+	push(@out,"\t.byte\t".join(',',@_)."\n");
+	}
+
 sub main'data_word
 	{
-	push(@out,"\t.long $_[0]\n");
+	push(@out,"\t.long\t".join(',',@_)."\n");
+	}
+
+sub main'align
+	{
+	my $val=$_[0],$p2,$i;
+	if ($main'aout) {
+		for ($p2=0;$val!=0;$val>>=1) { $p2++; }
+		$val=$p2-1;
+		$val.=",0x90";
+	}
+	push(@out,".align\t$val\n");
 	}
 
 # debug output functions: puts, putx, printf
@@ -559,7 +659,6 @@ sub main'picmeup
 		{
 		local($tmp)=<<___;
 #if (defined(ELF) || defined(SOL)) && defined(PIC)
-	.align	8
 	call	1f
 1:	popl	$regs{$dst}
 	addl	\$_GLOBAL_OFFSET_TABLE_+[.-1b],$regs{$dst}
@@ -572,13 +671,12 @@ ___
 		}
 	elsif ($main'pic && ($main'elf || $main'aout))
 		{
-		push(@out,"\t.align\t8\n");
 		&main'call(&main'label("PIC_me_up"));
 		&main'set_label("PIC_me_up");
 		&main'blindpop($dst);
-		&main'add($dst,"\$$under"."_GLOBAL_OFFSET_TABLE_+[.-".
+		&main'add($dst,"\$${under}_GLOBAL_OFFSET_TABLE_+[.-".
 				&main'label("PIC_me_up") . "]");
-		&main'mov($dst,&main'DWP($sym."\@GOT",$dst));
+		&main'mov($dst,&main'DWP($under.$sym."\@GOT",$dst));
 		}
 	else
 		{
@@ -587,3 +685,41 @@ ___
 	}
 
 sub main'blindpop { &out1("popl",@_); }
+
+sub main'initseg
+	{
+	local($f)=@_;
+	local($tmp);
+	if ($main'elf)
+		{
+		$tmp=<<___;
+.section	.init
+	call	$under$f
+	jmp	.Linitalign
+.align	$align
+.Linitalign:
+___
+		}
+	elsif ($main'coff)
+		{
+		$tmp=<<___;	# applies to both Cygwin and Mingw
+.section	.ctors
+.long	$under$f
+___
+		}
+	elsif ($main'aout)
+		{
+		local($ctor)="${under}_GLOBAL_\$I\$$f";
+		$tmp=".text\n";
+		$tmp.=".type	$ctor,\@function\n" if ($main'pic);
+		$tmp.=<<___;	# OpenBSD way...
+.globl	$ctor
+.align	2
+$ctor:
+	jmp	$under$f
+___
+		}
+	push(@out,$tmp) if ($tmp);
+	}
+
+1;
-- 
cgit v1.2.3-55-g6feb